AI Coding Workflow: Expansion, Not Just Speed
The discourse around AI coding tools is dominated by speed claims. "10x productivity!" But after synthesizing perspectives from Karpathy, Amodei, and Carmack, we think the real story is different.
The expansion thesis: The primary effect of AI coding agents isn't doing the same work faster — it's expanding what a single person attempts. You build things you never would have bothered with before. The scope of ambition per person grows dramatically.
Karpathy went from 80% manual to 80% agent coding in weeks, calling it "the biggest change in 2 decades." But he's clear-eyed about the limits: models behave like "hasty junior devs" — subtle conceptual errors, no pushback, tendency to overcomplicate.
The optimal workflow is declarative, not imperative. Give success criteria, not step-by-step instructions. The programmer becomes a specifier.
The tension: Amodei warns that expanding AI capability without understanding is dangerous. AI systems are "grown, not built" — emergent behavior means we can't predict or explain outputs. Interpretability is in a race against capability.
These perspectives are in direct tension. Karpathy celebrates expansion. Amodei warns that expanding blind is dangerous. Both are right.
Our position: Expansion with guardrails.
- Invest in specification skills — the bottleneck shifts from implementation to articulation
- Demand interpretability — don't treat AI as a black box you ship
- Maintain craft — atrophy is real, deliberately practice manual skills
- Embrace the expansion — the slopacolypse is real but so is the renaissance
AI coding is a power tool. Power tools build houses faster — but they also take off fingers. Learn the tool, respect the tool, keep your hands steady.
Sources: Karpathy (Jan 2026), Amodei "The Adolescence of Technology" (Jan 2026), Carmack #PaperADay (Jan 2026)